norskregnesentral/mccepy

Python package to generate counterfactuals using Monte Carlo sampling of realistic counterfactual explanations

mccepy's Issues

Loss of generality in postprocess-method for "higher_cardinality=True".

This problem was noticed while looking at examples/Examples_notebook_scratch.ipynb.

'postprocess' calls 'calculate_metrics', which later calls 'distance'. In the 'distance'-function, the data is inverse-transformed according to the 'inverse_transform'-function passed to the Dataset object, if the data is allowed higher cardinality than two levels per categorical feature. In the example, this object is constructed based on

class Dataset():
    def __init__(self, 
                 immutables, 
                 target,
                 categorical,
                 immutables_encoded,
                 continuous,
                 features,
                 encoder,
                 scaler,
                 inverse_transform,
                 ):
        
        self.immutables = immutables
        self.target = target
        self.feature_order = feature_order
        self.dtypes = dtypes
        self.categorical = categorical
        self.continuous = continuous
        self.features = self.categorical + self.continuous
        self.cols = self.features + [self.target]
        self.immutables_encoded = immutables_encoded
        self.encoder = encoder
        self.scaler = scaler
        self.inverse_transform = inverse_transform

where the inverse_transform function is defined as

def inverse_transform(df, 
                      scaler, 
                      encoder, 
                      continuous,
                      categorical,
                      categorical_encoded, 
                      ):

    df_categorical = pd.DataFrame(encoder.inverse_transform(df[categorical_encoded]), columns=categorical)
    df_continuous = pd.DataFrame(scaler.inverse_transform(df[continuous]), columns=continuous)

    return pd.concat([df_categorical, df_continuous], axis=1)

Now, the if-else statement in 'distance' fails if higher_card = True, i.e.

if higher_card:
        cf_inverse_transform = dataset.inverse_transform(cfs.copy())
        fact_inverse_transform = dataset.inverse_transform(fact.copy())

        cfs_categorical = cf_inverse_transform[categorical].sort_index().to_numpy()
        factual_categorical = fact_inverse_transform[categorical].sort_index().to_numpy()

    else:
        cfs_categorical = cfs[categorical_encoded].sort_index().to_numpy()
        factual_categorical = fact[categorical_encoded].sort_index().to_numpy()

because the inverse_transform-function misses four arguments. These can be passed quite easily however. Thus, I believe this problem can be fixed by changing the if-else statement above to

if higher_card:
        cf_inverse_transform = dataset.inverse_transform(cfs.copy(), dataset.scaler, dataset.encoder, continuous, categorical, categorical_encoded)
        fact_inverse_transform = dataset.inverse_transform(fact.copy(), dataset.scaler, dataset.encoder, continuous, categorical, categorical_encoded)

        cfs_categorical = cf_inverse_transform[categorical].sort_index().to_numpy()
        factual_categorical = fact_inverse_transform[categorical].sort_index().to_numpy()

    else:
        cfs_categorical = cfs[categorical_encoded].sort_index().to_numpy()
        factual_categorical = fact[categorical_encoded].sort_index().to_numpy()

I can send a PR later if this is interesting to repository owner.